Do not recompute _every_ entry matrix entries, but only

the ones with i<=j.  The rest follows from symmetry of
the Hessian.

Needless to say, this makes the Hessian assembly quite
a bit faster.

[[Imported from SVN: r8587]]